Bug Description

vision_analyze returns 400 invalid_request_error: "Invalid request data" for valid local image paths, across different primary chat models/providers.

This makes clipboard-pasted images and ordinary local files effectively unusable in the CLI for visual inspection workflows.

This appears to be a tool/integration issue rather than a model-specific limitation:

• reproduced with primary model anthropic/claude-opus-4-6
• reproduced again after switching the session to openai/gpt-5.4 via OpenRouter
• same failure shape in both cases

There is likely overlap with #6004 for clipboard images under ~/.hermes/images/, but the problem is broader: I also reproduced it on ordinary valid local JPEG/PNG paths outside the clipboard directory.

Steps to Reproduce

1. Start Hermes CLI on macOS.
2. Use vision_analyze on a valid local image path, for example:
   • ~/.hermes/images/clip_<timestamp>.png (clipboard image)
   • /path/to/project/runtime/artifacts/photos/example.jpg (ordinary local JPEG)
3. Ask for any simple description, e.g. "Describe the image in detail".
4. Observe tool failure.
5. Switch the primary session model/provider (I tested Anthropic -> OpenAI) and retry.
6. Observe the same failure again.

Expected Behavior

• vision_analyze should successfully analyze valid local images, or
• if local files are unsupported in the current execution path, it should return a clear actionable error (e.g. "local image paths are not mounted/available in sandbox") rather than a generic provider 400.

Actual Behavior

The tool returns a provider error like:

Error analyzing image: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'Invalid request data'}, 'request_id': 'req_...'}

This happens repeatedly on valid local images.

I also saw one size-related 400 on a larger upscaled JPEG:

messages.0.content.1.image.source.base64: image exceeds 5 MB maximum: 6937612 bytes > 5242880 bytes

That size error is expected for that specific file, but the main bug is that smaller valid images still fail with the generic Invalid request data 400.

Environment

• OS: macOS 15.3 / Darwin 25.3.0 arm64
• Repo: NousResearch/hermes-agent
• Version/Commit: 268ee6bd
• Python version: Python 3.9.6
• Primary chat models tested:
  • anthropic/claude-opus-4-6
  • openai/gpt-5.4 (via OpenRouter)

Error Output

Clipboard image path:

vision_analyze(image_url="~/.hermes/images/clip_<timestamp>.png")
→ Error code: 400
→ invalid_request_error: Invalid request data

Ordinary local JPEG path:

vision_analyze(image_url="/path/to/project/runtime/artifacts/photos/example.jpg")
→ Error code: 400
→ invalid_request_error: Invalid request data

Another local JPEG after switching primary model to GPT-5.4:

same 400 invalid_request_error: Invalid request data

Additional Context

• file identifies the JPEG as valid:

JPEG image data, JFIF standard 1.01, 1280x1197, components 3

• The clipboard-path failure may be partly explained by #6004 (clipboard images not mounted into Docker sandbox), but the ordinary local JPEG failure suggests the problem is not limited to ~/.hermes/images/.
• The repeated cross-provider failure suggests the bug is likely in the vision_analyze local-file handling / auxiliary vision request construction, not in the main chat model selection.
• A better error classification would help even before the root fix lands.